The regulatory trigger
Under the IRDAI Master Circular on Health Insurance Business (29 May 2024), cashless pre-authorisation is due within one hour of a complete request from the hospital, and final cashless authorisation at discharge within three hours. Reimbursement settlement is a separate, longer clock: a decision within thirty days of complete documents.
Incomplete packs pause the turnaround clock. Missing forms, unclear diagnosis notes, ambiguous cost estimates and KYC gaps are the usual causes. The breach is not that AI failed. It is that the pack was not ready when the clock was live.
How many cashless pre-auths last month waited on a pack that was already in the building but not yet readable?
The wedge
Cashless pre-authorisation document intake first. Not reimbursement settlement. Not medical adjudication. Analyse, identify, classify and map hospital packs (forms, estimates, prescriptions, ID, investigation reports) into the insurer or TPA schema. Per-field confidence and citation to the source region. Originals preserved. Every transform logged. Hindi and English mixed packs where that is the reality of the file.
Why the cloud answer fails here
Packs carry clinical detail, identity documents and policy data. Sending them to a public parsing API or a cloud agent workspace is a compliance and vendor-risk decision before it is a productivity decision. A metered cloud parser does not stop the one-hour clock. Residency on a public model is not the same as keys and weights the insurer controls.
What deployment looks like
Private AI Foundation installs once in the insurer or TPA environment: customer tenancy in-country, air-gapped, or bare metal. DocxIntel attaches as the intake application. Model Gateway, policy, DLP and audit sit on every document path. The customer holds the keys. Open weights (Apache 2.0 and MIT) sit on disk they control.
Operators keep queues and core systems. A human still authorises. AgentLab is not part of this first deal.
Controls in this workflow
These controls run on Private AI Foundation for every document ask. AgentLab, if used later, shares the same path.
- Packs stay inside the customer perimeter. Inference does not require a network path out.
- You hold the encryption keys in every tier that touches your data.
- Every document ask goes through the Model Gateway. There is no side door to a public model.
- Policy and DLP evaluate the request before weights run.
- RBAC names who may invoke which model class on which data class.
- DocxIntel writes per-field confidence and citation back to the source region. Originals are preserved.
- Consequential AgentLab actions wait for a named approver. Nothing silent-posts to your core systems.
- Immutable audit is written on the same install: who, what, which model, which policy, outcome.
- Apache 2.0 and MIT weights only. Licence text and an SBOM ship on disk per release.
What we do not claim
Not an IRDAI certification. Not a replacement for the claims core. Not automatic approval or denial. Not a free trial.
The paid proof of value
Thirty to forty-five days. Fixed fee. In the customer environment. One document type. Agreed real volume. Success criteria and conversion price agreed in writing before starting. If it misses the threshold, the customer keeps the report and owes nothing further.
This is not a free trial and not a sandbox that pretends to be residency. The cable can be pulled. The packs are yours.
